Investigation of an Ensemble Inflow-Prediction System for Upstream Reservoirs in Sai River, Japan

Abstract

In this study, an ensemble inflow-prediction system was developed for a hydropower-generation dam in the upper Sai River basin, and the accuracy of ensemble inflow prediction, which is important for efficient dam operation, was investigated. First, the Water and Energy Based Distributed Hydrological Model for Snow (WEB-DHM-S), a hydrological model developed for the Sai River basin, can represent the hydrological process from warm to cold seasons. Next, a system was developed on the Data Integration and Analysis System (DIAS) to predict inflows into the dam by inputting real-time meteorological data and ensemble rainfall forecast data into WEB-DHM-S. The WEB-DHM-S was calibrated and validated over a 3-year period from August 2015 to July 2018, and showed good agreement with observed inflows from base flow to peak flow and snowmelt runoff in each year. The results of inflow forecasting during frontal rainfall in August 2021 by inputting ensemble rainfall forecasts up to 39 h ahead showed that at the Inekoki Dam site, the total inflow (volume) to the peak was predicted with an accuracy of within 20% at 30 h, 24 h, 18 h, 12 h, and 6 h before the peak. These ensemble inflow forecasts can help optimize dam operations.

1. Introduction

Recently, the extreme events of heavy or no rainfall have become more frequent due to the effects of climate change. According to statistics based on AMeDAS observation data from the Japan Meteorological Agency (JMA), the number of heavy rainfall days (a daily rainfall of over 200 mm) has increased 1.5 times in the last decade, and the number of no rainfall days has also increased compared to the start of the observation period. “Climate Change in Japan 2020” [1], published by the Ministry of Education, Culture, Sports, Science and Technology (MEXT) and the Japan Meteorological Agency in December 2020, points out that the number of heavy rainfall days (over 200 mm/day) will increase, while the number of rainy days and the amount of snowfall and snowmelt will decrease, due to climate change. There are concerns about disasters due to heavy rainfall and drought events. In recognition of these situations, the Ministry of Land, Infrastructure, Transport and Tourism of Japan presented the “Dam Rehabilitation Vision”, a policy that states that, in addition to existing dams with flood-control functions, it is effective and necessary to utilize numerous hydropower dams across Japan by strengthening their flood-control function and improving their water-use efficiency. At the 4th Asia-Pacific Water Summit, held in Kumamoto in April 2022, the “Kumamoto Declaration [2]” indicated that efforts should be made for both climate change adaptation and mitigation measures. For example, focusing on floods, technologies should be developed both to reduce flood risks and use water resources effectively. On World Meteorological Day in March 2022, UN Secretary-General António Guterres announced that the UN will take the lead in taking new action to ensure that everyone on Earth has the protection of an early warning system within 5 years (Early Warnings for All), demonstrating the importance of flood forecasting [3]. In the case of droughts, it is essential to develop technologies for predicting inflow into dams during the dry season. They should also be capable of evaluating the amount of snow accumulated during the winter and the snowmelt discharge quantitatively while being deployable seamless and globally. Edwin et al. pointed out the importance of accurately estimating soil moisture, groundwater, and snow cover, which are used as the initial values for inflow forecasting [4]. It is critical to quantitatively assess inflows while calculating these initial values, which change every second. For these purposes, it is necessary to develop a reservoir inflow-prediction system that can predict river flow and dam lake levels in real time with high accuracy.

Inflow prediction has been studied by many researchers, for example, Ushiyama et al. [5] and Inomata et al. [6]. Seoro Lee et al. [7] investigated the issue using machine learning Auto-sklearn (AS). Real-time flood forecasting systems, such as GloFAS (Global Flood Awareness System) [8], Today’s Earth, and Today’s Japan [9], have been used for flood forecasting and have been proven effective. However, the hydrological models used in previous studies focus on flood events, but not on the effective use of “no rainfall” information.

In the case of mountainous basins, such as those with dams, kinematic wave approximations is possible, where the slopes, channel gradients, and water surface gradient can be considered the same. In addition, based on the fact that the flood travel time from the upstream area to the dam is determined by the slope and the distance of the river channel, it becomes possible to calculate areas with the same flood travel time at once, thus reducing the calculation time. With this idea, Yang et al. [10] developed a distributed runoff model named the Geomorphology-Based Hydrological Model (GBHM). Wang et al. [11] developed Hydro-SiB2, a one-dimensional model of water and energy response characteristics in the atmosphere and land surface, by coupling land surface processes and vertical infiltration processes in the soil layer. They also developed WEB-DHM [12], a water and energy balance-distributed hydrological cycle model. WEB-DHM divides the basin into several subbasins by Pfafstetter coding system [13] and then calculates the discharge for each flow interval (areas of the same flood arrival time), which contributes to reducing the processing time. Shrestha et al. [14] developed the Water and Energy Budget Based Distributed Hydrological Model for Snow (WEB-DHM-S), an advanced snowfall and snow-melting model. Naseer et al. [15] developed WEB-DHM-S in Oi river basin of Japan that that uses the output of an atmospheric model to give WEB-DHM-S a three-dimensional temperature distribution according to elevation within a basin to refine the temperature distribution for rain/snow discrimination of snowfall. WEB-DHM-S simulates hydropower efficiency and flood control by dividing a basin into a subbasin with a dam placed at its most downstream location. This is an effective model for inflow prediction.

Moiz et al. [16] validated the accuracy of inflow forecasting with a lead time of 3 months using WEB-DHM-S in the Kurobe River basin of the Kansai Electric Power Company, and showed the model’s effectiveness, especially during the snowmelt season. However, they did not investigate dam operations using real-time dam operation.

Koike et al. [17] applied the short-time (39 h) inflow-prediction and optimal dam-operation system to the Hatanagi I hydroelectric dam upstream of the Oi River and showed that the system can contribute to reducing flood risk and increasing hydropower generation. Nakamura et al. [18] investigated the short-time (39 h) and long-time (3 months) ensemble inflow-prediction and optimal dam-operation system and proved its effectiveness. Tamakawa et al. [19] validated the accuracy of real-time inflow prediction in the discharge event in 2018 in the Sai River, but indicated the need for further validation in other events.

The purpose of this study is to develop a Water and Energy Budget—Distributed Hydrological Model with Snow (WEB-DHM-S) for the upper Sai River basin in Nagano Prefecture, Japan, and develop a system that inputs atmospheric forcing and ensemble rainfall forecast data in real time to monitor and forecast inflows to the reservoir on the Data Integration and Analysis System (DIAS) [20,21]. The study also aims to validate the accuracy of the ensemble inflow prediction in the warm season at Nanakura and Inekoki Dams in the upper Sai River basin.

Since we are planning to study optimal operations for dams in the upper area, including flood control in the lower Sai River basin, this study focuses on the hydrological model development for the whole basin and verifies the inflow forecasting for the dams in the upper reach area.

2. Target Basin

The target basin is the Sai River basin in Japan. Figure 1 shows the locations of the basin. The Sai River has a basin area of 3054.5 km², with steep mountainous terrains represented by Yarigatake (3180 m), Hotakadake (3190 m), and Norikuradake (3026 m) in the upper reach and the flat Matsumoto Basin in the middle of the basin. The Sai River has a unique topography in the basin, with a riverbed slope of 1/20 in the mountains from its upper reach to Matsumoto City and a riverbed slope of 1/100 to 1/300 in the Matsumoto Basin located downstream, forming an alluvial fan area. There are 11 dams in the basin: Takase, Nanakura, and Omachi on the Takase River, Nagawado, Midono, and Inekoki on the Azusa River, and Ikusaka, Taira, Minochi, Sasadaira, and Odagiri on the lower Sai River. This study investigates ensemble inflow prediction at Nanakura and Inekoki Dams. The specifications of Nanakura and Inekoki Dams are given in

Table 1. Specifications of Inekoki and Nanakura Dams.

Data	Inekoki	Nanakura
Coordinates	36°10′06″ N 137°46′16″ E	36°29′25″ N 137°44′07″ E
Type of dam	Concrete Arch dam	Rock-fill dam
Impounds	Azusa river	Takase river
Height	60 m	125 m
Length	192.8 m	340 m
Total capacity	65,000 m3	7,380,000 m3
Surface area	51.0 ha	72 ha
Total storage capacity	10,700,000 m3	32,500,000 m3
Effective storage capacity	6,100,000 m3	16,200,000 m3
Construction year	1968	1979

.

3. Development of a Hydrological Model

To estimate the inflow to the dam with high accuracy and achieve effective dam operation for better flood control and more power generation, it is necessary to employ a hydrological model that can seamlessly and quantitatively estimate the base flow, the peak flow, and snowfall and snowmelt runoff. The model should also be able to correctly place dam locations within the watershed on the runoff model to reflect the effects of artificial dam operations.

This study used the the Water and Energy Budget Based Distributed Hydrological Model for Snow (WEB-DHM-S), developed and validated by Naseer et al. [15]. WEB-DHM-S was based on WEB-DHM, developed by Wang et al. [12]. It combines the land-atmosphere interaction model (Simple Biosphere model 2 (SiB2)), developed by Sellers et al. [22], and the distributed runoff model (GBHM), developed by Yang et al. [10]. WEB-DHM-S updates the land surface model to Hydro-SiB2 and improves the processes between the surface and soil layers and the snow process.

3.1. WEB-DHM

An overview of the basic model (WEB-DHM) is shown in Figure 2, citing Wang et al. [12] (pp. 3–4). The basin division from a basin to subbasins is based on Pfafstetter coding system (a), then subdivision from a subbasin to flow intervals comprising several model grids (b), then discretization from a model grid to a number of geometrically symmetrical hillslopes (c), and process descriptions of water moisture transfer from the atmosphere to river (d). Here, the SiB2 is used to describe the transfer of the turbulent fluxes (energy, water, and CO2 fluxes) between the atmosphere and land surface for each model grid, where Rsw and Rlw are downward solar radiation and longwave radiation, H is the sensible heat flux, and λ is the latent heat of vaporization. The GBHM simulates both surface and subsurface runoff using grid-hillslope discretization, and then simulates flow routing in the river network.

A distributed hydrological model starts with the collection of digital geographical information related to the study area for building a digital representation of the basin. A digital elevation model (DEM) is used to define the target catchment area, and the target basin is subdivided into subbasins (see Figure 2a). Within a given subbasin, a number of flow intervals are specified to represent the time lag and accumulation processes in the river network according to the distance to the outlet of the subbasin. Each flow interval includes a number of model grids (see Figure 2b).

Each model grid consists of a specific land use type and soil type; the SiB2 (Land Surface Model) is used to calculate turbulen fluxes between the atmosphere and land surface independently (see Figure 2b,d). Each model grid is subdivided into a number of geometrically symmetrical hillslopes (see Figure 2c), which are the basic hydrological unit (hliislope element) of WEB-DHM. For each hillslope element, DHM is used to simulate lateral water redistributions and calculate runoff (see Figure 2c,d). The runoff for a model grid is the total response of all hillslope elements in it.

For simplicity, the streams located in one flow interval are lumped into a single virtual channel in the shape of a trapezoid. All the flow intervals are linked by the river network generated from the DEM. All the runoff from the model grids in the given flow interval is accumulated into the virtual channel and let out to the outlet of the river basin.

WEB-DHM assumes that a large model grid comprises a set of symmetrical hillslopes located along the streams. Within a model grid, all hillslopes are viewed as being geometrically similar. A hillslope with unit width is a BHU and is represented by a rectangular inclined plane. The hillslope length within a model grid is calculated as l = A⁄(2ΣL), where A is the model grid area and ΣL is the total length of streams within the model grid extracted from the fine DEM. The total river length ΣL decreases with increasing threshold area. All streams extracted from the fine DEM within a given model grid can be simplified as one stream with a lenth ΣL (see Figure 3) flowing along the main flow direction of the model grid [12].

3.2. WEB-DHM-S

WEB-DHM-S is divided into four modules: (a) atmosphere-surface interaction, (b) vertical infiltration process, (c) slope runoff process, including groundwater runoff, and (d) river runoff process. Added to it is a three-layer snow model developed by Shrestha et al. [23] and a scheme for estimating albedo changes in the growth of snowpack particles in the runoff model, which inherits the three-layered energy balance-based snowmelt module of Simplified Simple Biosphere, version 3 (SSiB3; Sun and Xue) [24] and the albedo scheme of Biosphere-Atmosphere Transfer Scheme (BATS; Dickinson [25]). WEB-DHM-S can simulate the variability of snow density, snow depth, and snow water equivalent, liquid water and ice content, snow albedo, snow layer temperature, and thermal heat due to conduction in nine biomes by updating the land surface model to Hydro-SiB2 and improving the processes between the surface and soil layers and the snow process. WEB-DHM-S has a “snapshot” function that allows the system to be stopped and restarted at any given time and incorporates a dam-operation module. In addition, temperature profile data are implemented to determine the threshold of snow and rainfall. The lapse rate of temperature decrease with lapse late in elevation is estimated from the JRA55 temperature vertical profiles. Figure 4 shows an overview of Hydro-SiB2+Snow, an updated version of WEB-DHM’s atmosphere-surface interactive processes and vertical seepage processes [14].

3.3. Data Preparation and Datasets

The data preparation of WEB-DHM is mainly based on the digital geographical information related to the study area. It includes preparing digital maps of different types, such as elevation, land use, soil parameters, and geological conditions. First, the topography is simulated using a grid-based DEM, and then the watershed delineation is performed. After that, other digital maps are prepared based on the delineated watershed. The watershed is divided into sub-catchments using Pfastetter coding system [13].

The sub-catchments are composed of a number of flow intervals, including the basic hydrological units, and this process simplifies the river routing simulation. WEB-DHM requires precipitation and meteorological parameters, such as air temperature, wind speed, short-wave and long-wave radiation, specific humidity, air surface pressure, and cloud cover. Moreover, it needs Moderate Resolution Imaging Spectroradiometer (MODIS) Leaf Area Index (LAI) and Fraction of Photosynthetically Active Radiation (FPAR) satellite parameters and the soil and water parameters of various soil types from the global digital soil map (FAO). The schematic diagram of procedures for the preparation of input data is shown in Figure 5.

3.3.1. Basin Delineation Using NK-GIAS

The DEM used in this study is based on the 10-m resolution provided by the Geospatial Information Authority of Japan (GSI) [26]. The Nippon Koei Geographic Information Analysis System (NK-GIAS) version 1.8 [27] was used as a tool for basin delineation. NK-GIAS is a GIS software that specializes in hydrological analysis for its main functions, and is characterized by its compactness, easy operation, and quick response.

WEB-DHM-S performs basin division and creates river networks based on the concept of Pfafstetter coding system [13]. The basin is basically divided into nine subbasins, and if a subbasin is large as a subbasin, it is further divided into sub-subbasins. The flow is calculated by configuring cells called “flow interval”, which organizes a similar distance from the outlet of a subbasin. The model sometimes fails to create a river channel correctly when using DEM only. Such problems may arise in some cases: for example, when steep slopes, such as those of the Sai River, exist within a basin, when channels with different basin boundaries are located next to one another, or when flat topography exists widely within a basin.

The processing using DEM only creates an incorrect watershed boundary by combining different river channels at the constriction in a steep basin. To solve these problems, NK-GIAS implements the “River Channel Tool” using the DEM and the river channel model together to create the correct basin division and river channel even in narrowed areas.

Additionally, the system has a function to divide the basin based on Pfafstetter coding system that each dam is located at the most downstream point of a subbasin. This function enables the user to prepare an appropriate watershed that can effectively consider the effects of dam operations. In this study, the grid size of the Sai River basin was set to 250 m. Figure 6 shows images of the study basin with the steepest path for each cell and the flow intervals overlaid using NK-GIAS.

3.3.2. Static Data (Land Use and Soil Data)

The land use data was 1-km grid data from the U.S. Geological Survey (USGS) [28], and the soil data was 9-km spatial resolution data from the Food and Agriculture Organization of the United Nations (FAO) [29]. The vegetation in the land use data was reclassified and used according to Sib 2 classfication, and the soil type was set in the soil data and the corresponding hydraulic conductivity saturated soil water content of the surface; Van-Genutchen parameters like alpha and n are used as calibration parameters for the hydological model.

3.3.3. Dynamic Data (Meteorological Forcing Data)

To calculate the water and energy budget of the ground surface, data on precipitation, wind direction, wind speed, temperature, humidity, downward longwave radiation, and solar radiation are required. JMA radar AMeDAS analysis rainfall [30] was used for precipitation, along with JMA 55-year long-term re-analysis (JRA55), which is a re-analysis of data assimilated from the maximum available observations dataset using the latest numerical weather forecasting systems. Although it is a model output, it has an accuracy close to that of observations [31,32] and a spatial resolution of 0.5625°. A Temperature data with 3D information was used, with a spatial resolution of 1.25°, and a temporal resolution of every 6 h was used for wind direction, wind speed, humidity, downward longwave radiation, and solar radiation. Since the snowfall, snow accumulation, and snowmelt depend on temperature differences with elevation, the temperature data were generated using the method from Naseer et al. [15], which estimates the decrease in lapse rate with elevation increments from the vertical profile of JRA55 and extrapolates in three dimensions using AMeDAS ground-based temperature data in and around the Sai River basin.

3.3.4. Dynamic Data (Vegetation Data)

Leaf Area Index (LAI) and Fraction of Photosynthetically Active Radiation (FPAR) data with a spatial resolution of 1 km were used as vegetation data taken from TERRA MODIS 8-day composite data [33] provided by NASA. These were used to estimate evapotranspiration.

3.3.5. Discharge Data

Hourly discharge data at nine dams and two discharge stations in the basin provided by TEPCO Renewable Power, Incorporated (TEPCO Renewable Power) were used to calibrate and validate the WEB-DHM-S. The dataset used in WEB-DHM-S development is shown in

Table 2. Used dataset in the development of WEB-DHM-S.

Data	Spatial Resolution	Temporal Resolution	Data Source
Static Data
Digital Elevation Model	10 m	Fixed	GSI
Land use	1 km	Fixed	USGS
Soil Texture	9 km	Fixed	FAO
Dynamic Data (Meteorological Forcing Data)
Wind, humidity, downward longwave radiation, and solar radiation	0.5625 deg	6 h	JRA55
Air Temperature	1.25 deg	6 h	JRA55
LAI/FPAR	1 km	8 days	TERRA/MODIS
Precipitation	1 km	30 min	JMA Radar analysis rain
Discharge	Point	1 h	TEPCO

. These data were col-lected during the period 2017–2020 as part of a joint research project between the University of Tokyo, ICHARM, Nippon Koei, and TEPCO Renewable Power.

3.4. Calibration and Validation of WEB-DHM-S

3.4.1. Calibration

The Sai River basin has a snow area in the upper basin. Therefore, the model was calibrated for the warm season, in which there is no effect of snowmelt runoff, and then calibrated for the cold and snow melt season. Hydro-SiB2 was given default soil parameters and run to stabilize soil moisture for about 5 years, with the resulting vertical profiles of soil moisture as initial values. Then, soil-related parameters were adjusted from July–September 2018 and snow cover was calibrated from November 2017 to May 2018.

The period unaffected by snowmelt runoff was defined as the warm season (July–September), and the calibration was validated for July–September 2018 and the same months in 2015, 2016, and 2017. The results are shown in Figure 7. A NASH coefficient of 0.90 was obtained throughout the study period, with good results for the large discharges and low flows in July 2018, corresponding to the observed values. The Nash coefficients [32] were validated for the warm season of each year. They were 0.90, 0.86, and 0.94, respectively, indicating that floods and low flows can be well-represented.

Next, WEB-DHM-S was validated during the cold season (1 November 2017—30 June 2018), including the snowmelt season, and the snow cover area was calibrated by comparing the estimated snow-cover area with that observed by MODIS. The results are presented in Figure 8. The results show that the snow-cover area can be well-represented by WEB-DHM-S throughout the period.

The comparison of initial/default values and change of parameters for the calibration of soil and land are presented in

Table 3. The comparison of initial/default values and change of parameters for the calibration of soil and land.

Static Parameters	Symbol	Unit	Default	Optimized	Optimized
				Takase	Nagawado
Saturated hydraulic conductivity (surface)	ksat1	mm/h	23.68	47.36	47.36
Saturated hydraulic conductivity (root zone)	Ksat2	mm/h	2.368	2.368	2.368
Hydraulic conductivity of ground	Kg	mm/h	1.18	11.8	5.9
Residual soil moisture content	θr	vol/vol	0.074	0.074	0.074
Saturated soil moisture content	θs	vol/vol	0.47	0.564	0.564
alpha (Van-Genutchen Parameter)	alpha	-	0.014	0.017	0.017
n (Van-Genutchen Parameter)	n	-	1.49	2.235	2.682

.

3.4.2. Validation

Because the Sai River has a large basin area and a complex and diverse topography, the runoff model was calibrated separately for the upper and lower basin areas, respectively. Figure 9 presents the results of WEB-DHM-S validation at Takase and Nagawado Dams, the most upstream dams, for 3 consecutive years from August 2015 to July 2018. Figure 9 shows good agreement each year for the peak flow at Takase and Nagawado Dams in summer (July to October), the base flow in winter (November to March), and snowmelt runoff (April to June). The Nash–Sutcliffe coefficient (NSE) [34] defined in Equation (1) was used to validate the accuracy. The corresponding values are shown in

Table 4. The NSE values of individual dams in rom August 2015 to July 2018.

Dams/NSE	2015	2016	2017
Takase	0.53	0.63	0.55
Nagawado	0.70	0.78	0.79

.

The NSE values of individual dams in respective years are given in Table 4:

$$NSE=1-{\displaystyle \frac{\sum _{i=1}^N{\left(Q_{obs}-Q_{est}\right)}^2}{\sum _{i=1}^N{\left(Q_{obs}-\overline{Q_{obs}}\right)}^2}}$$

(1)

where $Q_{obs}$ and $Q_{est}$ are the observed and simulated discharges at the time interval and $\overline{Q_{obs}}$ is the averaged observe discharge. Note that the observed inflows at Takase and Ngawado Dams include negative values due to pumped-storage power generation. For details, see Tamakawa et al. [19].

4. Development of an Ensemble Inflow-Prediction System for Dams

The development policy for a real-time inflow-prediction system for dams is to use the latest data currently available with high temporal and spatial resolutions, and convert a complex set of data with different updating timings, and temporal and spatial resolutions, into a format that WEB-DHM-S can read and process. In this study, the system was developed on the Data Integration and Analysis System (DIAS), which is a platform for collecting, permanently accumulating, integrating, and analysing data from global and regional observations, integrating and fusing them with a hydrological model, updating them to a high level of information, and supporting problem-solving for decision-makers. It has 4 V (Volule, Variety, Velocity, and Veracity) capabilities and has an environment that enables real-time processing and execution of a wide variety of enormous data, as required in this study. A real-time 39-h-ahead ensemble inflow-prediction system was developed on DIAS using JRA55, JMA radar AMeDAS analysis rainfall, JMA short-time precipitation forecast, and regional ensemble forecast precipitation developed by Ushiyama et al. [35].

4.1. Regional Ensemble Rainfall Forecast Data

Regional ensemble forecast precipitation data were used by Ushiyama et al. [35]. They were calculated by combining the Weather Research and Forecasting (WRF) model developed at the National Center for Area Research (NCAR) and the Local Ensemble Transform Kalman Filter (LETKF). The JMA Global Spectral Model (GSM) forecasts, provided four times a day every 6 h, were used as boundary conditions, and the WRF model was used to calculate forecasts up to 9 h ahead for a 15-km mesh. Next, LETKF analysis was performed using global observations called PREPBUFR, collected and archived by the National Centers for Environmental Prediction (NCEP), to update the initial ensemble values. From the updated initial values, the WRF model was used to conduct forecasting again, and the process was repeated sequentially to continuously create an ensemble of initial values that reflected real-world weather conditions. This 9-h forecasting with 32 ensemble members was extended, and the inner 3-km mesh domain was also calculated to obtain high-resolution forecast data for the next 48 h. For real-time operations, the first 9 h of values were not used, but data up to 39 h ahead were used, considering the time required for ensemble calculations and the time required to mitigate the effects of the initial forecasting conditions.

4.2. Real-Time Ensemble Inflow-Prediction System

The real-time prediction of inflow into a dam using WEB-DHM-S consists of two steps: (1) calculation of the amount of surface and soil conditions (e.g., soil moisture content, groundwater level) and the amount of inflow up to the “current time”, and (2) ensemble prediction of inflow using rainfall forecast data from the “current time”. A schematic image of the real-time ensemble inflow-prediction system is shown in Figure 10.

4.2.1. Calculation of the Amount of Surface and Soil Conditions Up to the “Current Time”

The JRA55, JMA radar AMeDAS analysis rainfall, and forecast data from Ushiyama et al. were used to calculate the amount of state up to the current time. JRA55 data up to 3 days before the “current time” was provided to DIAS. JMA radar AMeDAS analysis rainfall was updated hourly and provided to DIAS in real time. Therefore, up to 3 days before the current time, atmospheric forcing data from JRA55 was used to calculate the amount of inflow and the surface and soil conditions. For the remaining 2 days up to the present time, precipitation was calculated using JMA radar AMeDAS analysis rainfall, and other atmospheric forcing was calculated using data from Ushiyama et al.

4.2.2. Calculation of Inflow Prediction from “Current Time”

For the prediction from the “current time,” JMA Short-Time Forecast and the ensemble forecast data from Ushiyama et al. were used. In this study, JMA Precipitation Short-Time Forecasts, which were provided to DIAS every 10 min in real time, were used up to 6 h ahead. For the next 33 h, the ensemble prediction updated every 6 h by Ushiyama et al. was used.

5. Ensemble Inflow Hindcast Simulation at Nanakura and Inekoki Dams

The 32 ensemble, 39-h-ahead inflow hindcasting at two dams, Nanakura (Takase and Nanakura Dam subbasin) and Inekoki (Nagawado, Midono, and Inekoki Dam subbasin), were conducted for a large discharge event in August 2021. Both dams are classified as “Class 1 (a dam that requires flood storage to deal with flooding).” The results of the large discharge are shown in Figure 11. In Figure 11 the black line shows the observed inflow, and the light blue line shows the inflow 39 h ahead based on the 32 ensemble averages. The results show that the timing of the increase of the peak and the decrease of the discharge are well-represented. The results of this study are based on the application of feedback (i.e., the predicted results are shifted to the first inflow at the time of the observation in question).

In the case of a Class I dam, it is possible to store inflow according to the availability of the reservoir, so it is important to predict the total inflow (volume) up to the peak, rather than the time of peak arrival. The ratio of the total predicted inflow volume up to 39 h before the peak flow occurrence time (30, 24, 18, 12, and 6 h) to the total observed inflow volume during the same time period was evaluated by setting the accuracy index of 20%. The results are shown in Figure 12 and

Table 5. The ratio of the total predicted inflow volume up to 30 h before the peak flow occurrence time (30, 24, 18, 12, and 6 h) to the total observed inflow volume during the same time period].

Dams	Peak Discharge (m3/s)	Peak Date and Time	30 h	24 h	18 h	12 h	6 h
Inekoki	651.0	14 August 2021 22:00	0.88	0.95	0.83	1.05	1.02
Nanakura	271.5	15 August 2021 04:00	1.25	1.19	1.58	1.32	0.94

.

In the case of Inekoki Dam, the predicted inflows showed a high accuracy of within 20% from 30 h in advance. Nanakura Dam also showed an accuracy of within 20%.

6. Result and Discussion

This study investigated a system designed for a power-generation dam in the upper Sai River in order to integrate satellite observation data, ground observation data, and numerical weather prediction model outputs (reanalysis data) and predict inflow in real time. We calculated the distribution of snow cover and soil moisture on a basin scale at a high temporal and spatial resolution of 250 m at hourly intervals using a hydrological model that takes into account snowfall, snow cover, and snowmelt on a basin scale, which cannot be observed on a high spatiotemporal scale. We then tested inflow monitoring and prediction using the calculation results as the initial conditions and found that our proposed system can plausibly predict dam inflow in real time by integrating different types of data.

This system requires an environment that allows users to archive, process, and analyze instantaneously, and visualize and share a huge amount of different data, including domestic and global satellite and ground observations and numerical weather model outputs. However, we confirmed that the system can run reliably on the Data Integration and Analysis System (DIAS) operated by the University of Tokyo under the Global Environmental Data Integration and Analysis Platform Project of the Ministry of Education, Culture, Sports, Science and Technology of Japan.

Still, the uncertainty of the input rainfall must be considered in applying this system. Users need to pay attention to the range of ensemble forecasts when optimizing dam operations based on inflow forecasts. The main advantage of this system is that it is applicable anywhere in the world where discharge observation data are available for validation. In addition to better flood control, the system can help improve hydroelectric dam operations for more power generation by predicting snowmelt runoff. Moreover, by inputting long-term weather forecast data as atmospheric forcing and weather forecast data considering climate change impacts, the system can predict dam inflow based on future snowfall, snow cover, and snowmelt within the uncertainty of the respective data. Studying hydropower dam operations for more efficient electricity generation based on such future projections could also make a significant contribution to planning climate change adaptation measures and carbon neutral policies in national and international organizations.

Our system also enables real-time monitoring of snow. Despite its usefulness as an energy, water, and tourism resource, snowfall has been a challenge for real-time quantitative measurement in mountain basins, cryospheres, and other locations. The proposed system can be a practical solution to this long-term problem.

7. Summary

In this study, an ensemble inflow-prediction system was developed for a power-generation dam in the upper Sai River basin, and the accuracy of the ensemble inflow prediction, which is important for efficient dam operation, was investigated.

First, an overview of the WEB-DHM-S was introduced, together with details of the basin delenation data for input (Static Data), atmospheric forcing data for driving the model (Dynamic Data), and inflow data for validation. The calibration and validation of WEB-DHM-S was carried out over a 3-year period from August 2015 to July 2018 and showed good agreement with observed inflows from low to high flow and snowmelt flow in each year. Next, an overview of ensemble rainfall forecast data was presented and a system was built to predict inflows to dams in the upper reach by inputting meteorological data and ensemble rainfall forecast data in real-time into WEB-DHM-S on the Data Integration and Analysis System (DIAS).

The results of inflow forecasting during frontal rainfall in August 2021 by inputting ensemble rainfall forecasts up to 39 h ahead showed that at the Inekoki Dam site, the total inflow (volume) to the peak was predicted with an accuracy of within 20% at 30 h, 24 h, 18 h, 12 h, and 6 h before the peak, respectively. This has led to results that lead to the investigation of optimal dam operations based on ensemble inflow predictions. And the prediction accuracy at Nanakura Dam was within 20% at 24 h and 6 h before the event.

It showed that this inflow-forecasting result can be used to investigate the optimal dam operations for effective power generation, including downstream flood control and pre-release of water from the hydropower-generation dam. In the future, the entire basin, including the lower reaches of the Sai River, will be considered for optimal dam operations, including more power generation and better flood control in the lower Sai River basin.